Human Feedback

Human feedback is the raw material that turns a language model into an assistant with behavioral values. It might take the form of raters picking the better of two responses, writers correcting a model’s mistakes, or users clicking a thumbs-up or thumbs-down button in a product. Each signal, aggregated across many examples, shapes what the model learns to do. The challenge is that human feedback reflects the perspectives, assumptions, and blind spots of whoever provided it — so who gives feedback, under what conditions, and with what guidance all matter enormously. For behavior architects, this is a core reason to think carefully about annotator selection, instructions, and quality control.